The scaffold made the difference

With Qwen3.5-9B Q4 weights left unchanged and only the scaffold swapped, results on the Aider Polyglot 225 questions changed dramatically.

• vanilla Aider: 19.11%
• little-coder: 45.56% average pass@2
• The little-coder figure is the average of two full runs.

The key factor was not swapping the model but scaffold–model fit. The author tuned the scaffold to match the behavioral characteristics of a small local model.

• Limited reasoning budget
• A Write guard that doesn't overwrite existing files
• Explicit workspace exploration
• Injecting small skills per turn instead of one large fixed preamble

This isn't validation at the level of a formal paper, but the effect size was large enough to be considered a signal worth sharing. What's still needed are more repeated trials, component ablations, other model families, and other benchmarks.